[E2E Testing] KV-Cache (#1004)

~~Contingent on merge of
vllm-project/vllm#11354
^ merged

SUMMARY:
Add kv-cache e2e testing
* One small model - tinyllama - with kv-cache
* One small model - tinyllama - with kv-cache + gptq
* Fused Model - phi3 - with kv-cache

Signed-off-by: Kyle Sayers <[email protected]>

Author: horheynm

tests/e2e/vLLM/configs/kv_cache_gptq_tinyllama.yaml

@@ -0,0 +1,7 @@
+cadence: "nightly"
+test_type: "regression"
+model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
+recipe: tests/e2e/vLLM/recipes/kv_cache/gptq.yaml
+dataset_id: HuggingFaceH4/ultrachat_200k
+dataset_split: train_sft
+scheme: kv_cache_default_tinyllama

tests/e2e/vLLM/configs/kv_cache_phi3.yaml

@@ -0,0 +1,7 @@
+cadence: "nightly"
+test_type: "regression"
+model: microsoft/Phi-3-mini-4k-instruct
+recipe: tests/e2e/vLLM/recipes/kv_cache/default.yaml
+dataset_id: HuggingFaceH4/ultrachat_200k
+dataset_split: train_sft
+scheme: kv_cache_default_phi3

tests/e2e/vLLM/configs/kv_cache_tinyllama.yaml

@@ -0,0 +1,7 @@
+cadence: "nightly"
+test_type: "regression"
+model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
+recipe: tests/e2e/vLLM/recipes/kv_cache/default.yaml
+dataset_id: HuggingFaceH4/ultrachat_200k
+dataset_split: train_sft
+scheme: kv_cache_default_tinyllama

tests/e2e/vLLM/recipes/kv_cache/default.yaml

@@ -0,0 +1,5 @@
+quant_stage:
+  quant_modifiers:
+    QuantizationModifier:
+      kv_cache_scheme:
+        {num_bits: 8, type: float, symmetric: true, strategy: tensor}

tests/e2e/vLLM/recipes/kv_cache/gptq.yaml

@@ -0,0 +1,17 @@
+quant_stage:
+  quant_modifiers:
+    QuantizationModifier:
+      kv_cache_scheme:
+        {num_bits: 8, type: float, symmetric: true, strategy: tensor}
+    GPTQModifier:
+      sequential_update: false
+      ignore: ["lm_head"]
+      config_groups:
+          group_0:
+              weights:
+                  num_bits: 4
+                  type: "int"
+                  symmetric: true
+                  strategy: "channel"
+                  actorder: False
+              targets: ["Linear"]